inside
The employee magazine
of the Volkswagen brand

Digitalization

The car that’s learning to learn

Experts all agree: Cars will no longer need drivers. Before that happens, Volkswagen needs to teach them how to learn.

Driverless cars will be part of our future. But that will take a few years, and there will also be various challenges along the way. Because driving is a complex matter: braking, accelerating, judging speeds – drivers have to make a multitude of decisions every second. It’s good that people are remarkably adaptive, and that their skills become more practiced with every mile on the road. Our minds use experience to develop certain behavioral patterns, and we learn to respond faster to driving situations because we have gone
through them or similar ones in the past.

When a computer takes the wheel, the main question that fires up researchers and developers is how we can teach it to learn, so it will continuously improve its response to traffic conditions. The magic term here is “machine learning.” Florian Neukart from Volkswagen Data:Lab works on this topic
in particular, “For driverless cars, we need software that teaches itself how to grow and change.” And this is how it works: “Whenever the computer gets new data from its
environment, let’s say from camera images, it searches for patterns and adjusts its behavior. Or in other words, it learns.”

Cars of the future need to do what people can do - learn to learn.

That might sound simple, but it’s actually quite tricky. One example: programmers have developed automobile software that should make the car brake when a pedestrian enters the field of view. That has to work with every pedestrian in the world – large, small, fat, thin, coming from the right or the left. Only, no programmer can show his software pictures of any pedestrian in the world to teach it who it has to brake for. “This is where machine learning starts,” says Neukart. “The computer is trained to recognize a person, by means of certain features such as legs, arms, or types of movement.”

“This is where machine learning starts“, says Neukart. “The computer is trained to recognize a person, by means of certain features such as legs, arms, or types of movement.” The more the computer knows, the easier the recognition process. And because it can infer from all the properties that make up a human, the computer doesn’t care whether the pedestrian has a beard or half the body is concealed behind a trash can, before he or she crosses the road. The computer then registers: Warning, person! – and brakes. Experts hope that when cars are equipped with this type of intelligence, the number of traffic incidents will be considerably reduced.